MemTensor Open-Sources MemT-1.0, a Memory-Native Foundation Model

On October 20, the Memory Tensor team released and open-sourced MemT-1.0, a "memory-native" foundation model that makes retaining, updating, and recalling experience a learnable capability of the model itself — rather than relying on external memory stores or repeatedly injecting context. It is being called the industry's first memory-native foundation model and the largest open-source parametric memory model.

Confirmed

• Architecturally, MemT-1.0 is built on the Metis backbone with Gated DeltaNet-related designs, introducing a continuously updatable internal memory state alongside the Transformer trunk so that parametric memory directly participates in generation (m1, m5)
• The backbone carries general capabilities, while the dynamic memory state serves as a dedicated working state independent of users/tasks, updated at the end of each conversation turn; the team claims the model can recall details from 500 turns of dialogue "closed-book" (m4)
• The family includes multiple variants: MemT-1.0-Nano has 36B parameters with 3B activated, built on the Qwen3.6-35B-A3B base; the largest variant reaches 288B parameters (m3)
• Performance: 1.86x faster inference at 128K context (m3); SWE-bench success rate rose from 54% to 61.6%, a gain of over 3 points (m2, m5)
• The model has been open-sourced (m3, m5)

Why it matters

• 赛博禅心 pointed out three problems with external text-based memory: retrieval rules must be pre-defined, context grows longer and costlier with use, and true reuse is impossible; earlier approaches that trained history into parameters (such as Temp-LoRA) were prohibitively expensive. MemT-1.0 targets exactly this pain point, training the model to learn to read, write, update, and invoke memory on its own (m2, m4)
• If the parametric memory approach proves viable, the cost and reusability problems of long conversations and experience accumulation could be fundamentally alleviated — an architecture-level challenge to the current RAG/context-injection paradigm

2026-10-20 ~ 2026-10-20 · 5 related posts

• MemTensor's MemT1.0 lets models learn to read and write their own memory — APPSO · 2026-10-20
• [source] MemTensor open-sources MemT1.0, a 288B memory-native model with 1.86x faster 128K inference — 量子位 · 2026-10-20
• MemT1.0 internalizes memory into model weights, lifting SWE-bench from 54.0% to 61.6% — 赛博禅心 · 2026-10-20
• [source] MemTensor releases MemT-1.0, a memory-native architecture embedding continuous memory into model computation — MemTensor · 2026-10-20

1 near-duplicate retellings: 机器之心